Auto-compaction runs repeatedly with no user-facing notification, before the context meter reaches 100%

Summary — During a long Claude Code session, the conversation was auto-compacted at least ~3 times with no visible notification. My context-usage indicator never reached 100%, yet compaction happened repeatedly, and there was no on-screen indication any of it occurred. I only found out when Claude grepped the session transcript and found the compaction markers after the fact.

Environment — Claude Code desktop app; Windows 11; long multi-hour session.

What happened — The session JSONL transcript contains isCompactSummary + PostCompact markers at ~3 distinct points (auto-compaction events). None produced a user-facing notification. The context-usage indicator visibly dropped at these points (expected, since the window is replaced by a shorter summary) but with no explanation — which reads as silently lost context.

Expected — A clear, visible notification each time auto-compaction runs (ideally a brief warning before it fires, with the option to intervene / /compact manually), and a context indicator that makes it obvious a compaction occurred rather than just silently dropping.

Impact — With no notification, there's no way to know context was summarized, or whether anything important was dropped, without manually inspecting the transcript. It erodes trust in long sessions.

Note — This is about the harness auto-compaction + its notification UX, not model output.